program synthesis

François Chollet argues that scaling deep learning alone will not deliver human-like general intelligence, which he defines as learning new tasks with human-level efficiency. At Indium, his team is exploring program synthesis and symbolic descent to find compact models that learn from less data and generalize better. He explains how verifiable rewards and post-training have driven progress in coding agents and ARC-AGI benchmarks, while warning that benchmark saturation can reflect engineering effort rather than greater fluid intelligence. ARC-AGI-3 tests whether agents can explore unfamiliar interactive environments efficiently, and Chollet predicts AGI could emerge around 2030. He also discusses building self-improving AI research, lessons from creating Keras, and how expertise can help people use AI progress to their advantage.

François Chollet argues that scaling pretraining produces useful task-specific skills but does not, by itself, create fluid intelligence: the ability to adapt to unfamiliar situations. He describes intelligence as efficiently reusing past experience to handle novelty, and explains why benchmarks based on familiar tasks can reward memorization instead. ARC-1 and ARC-2 probe different limits of AI adaptation, while ARC-3 is intended to test interactive agency. Chollet proposes combining learned intuition with discrete program search, so AI can assemble solutions from reusable abstractions, and says his new research lab AIndy is building toward that approach.